README.md

metadata

license: cc0-1.0
task_categories:
  - audio-classification
tags:
  - audio
  - audio-generation
  - fsd50k
  - rnn
  - music
  - ambient
size_categories:
  - 1K<n<10K

FSD50K CC0 Curated v1

A 1,408-clip CC0-only subset of FSD50K (Fonseca et al., 2022), curated for an RNN/LSTM audio generation teaching assignment.

Contents

• 1,408 WAV files from the FSD50K dev split (<file_id>.wav)
• fsd50k_cc0_dev_curated_v1_manifest.csv — per-clip metadata
• All files are CC0 / public domain — no attribution required
• 18 primary labels covering music instruments and nature ambient sounds
• Total size: ~1.5 GB, total duration: ~4.73 hours
• Original sample rates and durations preserved (no resampling or truncation)

Manifest columns

Column	Description
file_id	FSD50K numeric ID — matches WAV filename
wav_path	Original local path (ignore — use file_id to locate files)
split	Train / val / test assignment (stratified by primary_label, seed=42)
license_raw	Original Freesound license URL
license_normalized	Always CC0 in this subset
labels	Comma-separated FSD50K labels
primary_label	Single label used for stratification (first keep-label match)
tags	Freesound user tags
size_bytes	WAV file size in bytes
duration_sec	Clip duration in seconds

Curation

Clips were selected from the 14,959 CC0 dev files using:

1. Exclude-label filter (speech, vehicles, household noise, etc.)
2. Keep-label requirement (music + nature ambient categories)
3. Duration threshold ≥ 4.17 s (with 6 anchor clip exceptions)
4. Round-robin category-balanced fill to a 1.5 GB ceiling

Seed: 42. Full details in the source repository's doc/DATASET_CURATION_REPORT.md.

Usage

from huggingface_hub import snapshot_download
snapshot_download(repo_id="hughxuechen/fsd50k-cc0-curated-v1", repo_type="dataset", local_dir="data/fsd50k_preprocessed")

Citation

Fonseca, E., Favory, X., Pons, J., Font, F., & Serra, X. (2022). FSD50K: An Open Dataset of Human-Labeled Sound Events. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30, 829–852.